A retrospective cohort study using machine learning to predict coronary artery lesions in children with Kawasaki disease.

BACKGROUND

Kawasaki disease (KD) mainly occurs in children under 5 years old, and the most common complication of KD is coronary artery lesion (CAL). In recent years, the incidence rate of KD has increased year by year worldwide, so it is particularly important to strengthen the diagnosis of KD and identify CAL early.

METHOD

This retrospective cohort study included a total of 436 children diagnosed with Kawasaki disease and aimed to develop a predictive model for CAL using early clinical symptoms and laboratory features. To reduce potential confounding, propensity score matching (PSM) was applied, and both univariate and multivariate analyses were conducted to identify significant predictors of CAL. Subsequently, through machine learning, a predictive column chart model was constructed using clinical features and routine laboratory blood indicators, and the model was evaluated using ROC curves, calibration curves, and DCA curves.

RESULT

This study found that gender, medical history, cough, diarrhea symptoms, and high CRP levels were independent risk factors for concurrent CAL. To further predict CAL risk, a column chart model was constructed based on LASSO regression and ten fold cross validation. The ROC curve in the training queue showed good discriminative ability (AUC: 0.879), while the ROC curve in the validation queue showed good discriminative ability (AUC: 0.859). This model exhibits good discriminative ability, high accuracy, and potential clinical benefits in both training and validation sets.

CONCLUSION

Through this study, we provide clinicians with a new tool to more accurately predict and manage CAL risk in children with KD, which can help optimize treatment strategies and improve efficacy.